This repository contains the code of Medium-Difficulty Samples and Logit Reshaping (MDSLR) accepted at ICLR'25.

Part of the code is modified from CRD, PEFD and InfoBatch.

Python==3.6, pytorch==1.8.0, torchvision==0.2.1

You need to manually download ImageNet dataset and save it in './data'.

sh scripts/run_pretrained_teachers.sh

sh scripts/generate_random_index.sh

sh scripts/generate_medium_index.sh

sh scripts/run_cifar.sh

@inproceedings{
chen2025mediumdifficulty,
title={Medium-Difficulty Samples Constitute Smoothed Decision Boundary for Knowledge Distillation on Pruned Datasets},
author={Yudong Chen and Xuwei Xu and Frank de Hoog and Jiajun Liu and Sen Wang},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025}
}